# Learning What to Activate: Combinatorial Capability Allocation for Long-Horizon Multimodal Agents

arXiv: https://arxiv.org/abs/2609.27869 · HTML: https://mlsanity.com/paper/2609.27869/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 26 of its 43 references that are in the map.
- First posted on arXiv 20 Aug 2026.

## Builds on (most cited first, 26 of 26)

- [ReAct: Synergizing Reasoning and Acting in Language Models](https://mlsanity.com/paper/2210.03629/) (2022)
- [Toolformer: Language Models Can Teach Themselves to Use Tools](https://mlsanity.com/paper/2302.04761/) (2023)
- [OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments](https://mlsanity.com/paper/2404.07972/) (2024)
- [GAIA: a benchmark for General AI Assistants](https://mlsanity.com/paper/2311.12983/) (2023)
- [MMMU-Pro: A More Robust Multi-discipline Multimodal Understanding Benchmark](https://mlsanity.com/paper/2409.02813/) (2024)
- [ToolACE: Winning the Points of LLM Function Calling](https://mlsanity.com/paper/2409.00920/) (2024)
- [Multi-Agent Collaboration via Evolving Orchestration](https://mlsanity.com/paper/2505.19591/) (2025)
- [Agentic Reasoning: A Streamlined Framework for Enhancing LLM Reasoning with Agentic Tools](https://mlsanity.com/paper/2502.04644/) (2025)
- [A Unified Approach to Routing and Cascading for LLMs](https://mlsanity.com/paper/2410.10347/) (2024)
- [Organizing, Orchestrating, and Benchmarking Agent Skills at Ecosystem Scale](https://mlsanity.com/paper/2603.02176/) (2026)
- [BEST-Route: Adaptive LLM Routing with Test-Time Optimal Compute](https://mlsanity.com/paper/2506.22716/) (2025)
- [MMInA: Benchmarking Multihop Multimodal Internet Agents](https://mlsanity.com/paper/2404.09992/) (2024)
- [AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration](https://mlsanity.com/paper/2602.03786/) (2026)
- [AgentRouter: A Knowledge-Graph-Guided LLM Router for Collaborative Multi-Agent Question Answering](https://mlsanity.com/paper/2510.05445/) (2025)
- [OPID: On-Policy Skill Distillation for Agentic Reinforcement Learning](https://mlsanity.com/paper/2606.26790/) (2026)
- [SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning](https://mlsanity.com/paper/2607.14777/) (2026)
- [Adaptive Tool Use in Large Language Models with Meta-Cognition Trigger](https://mlsanity.com/paper/2502.12961/) (2025)
- [Beyond Bradley-Terry Models: A General Preference Model for Language Model Alignment](https://mlsanity.com/paper/2410.02197/) (2024)
- [Tool Preferences in Agentic LLMs are Unreliable](https://mlsanity.com/paper/2505.18135/) (2025)
- [EvoRoute: Experience-Driven Self-Routing LLM Agent Systems](https://mlsanity.com/paper/2601.02695/) (2026)
- [The Evolution of Tool Use in LLM Agents: From Single-Tool Call to Multi-Tool Orchestration](https://mlsanity.com/paper/2603.22862/) (2026)
- [MTRouter: Cost-Aware Multi-Turn LLM Routing with History-Model Joint Embeddings](https://mlsanity.com/paper/2604.23530/) (2026)
- [ACE-Router: Generalizing History-Aware Routing from MCP Tools to the Agent Web](https://mlsanity.com/paper/2601.08276/) (2026)
- [Routing with Generated Data: Annotation-Free LLM Skill Estimation and Expert Selection](https://mlsanity.com/paper/2601.09692/) (2026)
- [NaviAgent: Graph-Driven Bilevel Planning for Scalable Tool Orchestration](https://mlsanity.com/paper/2506.19500/) (2025)
- [Multi-Agent Routing as Set-Valued Prediction: A WildChat Benchmark and Cost-Aware Evaluation](https://mlsanity.com/paper/2606.28925/) (2026)

## Abstract (arXiv)

Long-horizon multimodal agents rely on specialized capabilities for perception, retrieval, reasoning, verification, and execution. Existing designs typically activate a fixed capability set or invoke a predefined workflow, incurring substantial computational overhead while failing to accommodate stage-dependent capability demands. In this paper, we study the textitcombinatorial capability allocation problem for long-horizon multimodal agent systems, where the system selects a cost-sensitive subset of specialized capabilities at each interaction stage, which is nontrivial since capability values depend on the selected subset, while previous allocations alter the states encountered by subsequent decisions. We introduce textscCoCA, an on-policy learning framework that recovers a deployable capability-subset policy from sparse conditional comparisons. On states visited by the student policy, the stronger teacher compares the marginal net values of candidate capabilities, conditioned on the currently selected subset. Then, we adopt a conditional utility model to transform such comparisons into an autoregressive capability-subset policy, avoiding explicit enumeration. We further introduce dual-level on-policy distillation to address distribution mismatch both across environment states and within the partial subsets encountered during set construction. Finally, trajectory-level reinforcement learning refines the distilled policy toward task success, activation cost, and allocation stability. At inference time, allocation is performed solely by the lightweight student policy without teacher queries or online updates. Experiments on long-horizon multimodal environments and controlled capability-demand shifts demonstrate the superiority of our method over the state-of-the-art baseline methods.